Will my photo get flagged as sensitive on X?
A plain-English guide to X's sensitive media labels, and how Will It Flag? estimates the answer on your own device before you post.
What "sensitive media" means on X
X (formerly Twitter) lets people share almost any kind of image, but some of it is placed behind a warning. When a post carries a sensitive media label, people who haven't opted in to seeing that kind of media get a blurred box with a "Content warning" message and a Show button instead of the picture. Under X's published rules, the label is used mainly for two broad groups of images: adult content, and graphic content such as visible injuries, gore or serious violence. Some material isn't allowed at all, label or not, and a label doesn't make it acceptable. X's own sensitive media policy is the source of truth and changes over time, so read it if you're unsure.
A label can be applied in three ways. You can mark your own media as sensitive in your privacy settings. X's automated systems can label a post when it's uploaded. And a post can be labeled after people report it and it gets reviewed. If an account repeatedly posts unlabeled sensitive media, X may set the label on the whole account. A label also affects reach: labeled media is hidden from people who haven't chosen to see it, and it can't be used in places like your profile photo or header image.
Why check before you post?
- Artists and photographers often find that figure studies, fashion shoots, beach photos or medical illustrations end up behind a warning, which can quietly cut engagement.
- Journalists and researchers sharing news imagery may want to label graphic photos themselves rather than risk an account-level label.
- Brands and community managers want to know whether a campaign image could be hidden from part of the audience before it goes live.
- Everyone else just wants to avoid surprises, like a gym selfie or a horror-movie screenshot getting blurred.
Will It Flag? gives you a quick, private second opinion. If a photo comes back as likely to be flagged, you can crop it, choose a different one, or mark it as sensitive yourself so viewers know what to expect.
How the check works
In 2023 Twitter published a large part of its recommendation and safety code on GitHub. It included the training code for its sensitive-media image classifier. That code describes a simple, well-known design: a pretrained image model turns a photo into a list of numbers (an embedding), and a small neural network on top turns those numbers into a probability that the image is sensitive. The thresholds that turn the probability into a yes/no decision are chosen from precision and recall curves, trading missed images against false alarms.
What X did not publish is the image model it uses, the trained weights, or its training data. So Will It Flag? rebuilds the published design from open parts:
- Image embedding. We use OpenAI's open CLIP image encoder (a smaller "Quick" version and a larger "Accurate" version), reduced to 256 numbers to match the input size in X's code.
- X's classifier head. We train the same small network X describes on public CLIP embeddings released by the LAION research group. No explicit images were collected or viewed to build this; only precomputed numeric embeddings and their labels were used.
- Thresholds. Like X's code, we pick a "borderline" and a "likely flagged" threshold from precision/recall curves on held-out data.
- Graphic violence add-on. X's open code covers only the sensitive-image model, so violence and gore are checked by a separate zero-shot CLIP test that compares your photo with text descriptions. It is clearly marked as an add-on.
- Cross-check. LAION's open detector runs alongside for comparison.
The result shows one overall verdict (Likely fine, Borderline or Likely flagged), a preview of how your post might look with a warning, and the individual scores so you can see why.
Your photo never leaves your device
All of this runs inside your browser, using WebGPU on your graphics chip when available, or WebAssembly on the processor otherwise. The first check takes a moment while the checker gets ready; after that, checks are fast. There is no upload step, no account and no server that receives photos. We don't store them, and we couldn't if we wanted to.
This site is free and paid for by ads. Ads come from third-party scripts, so we built in extra protections: the photo is shown only inside a sealed, sandboxed preview frame that other scripts on the page can't read; file selections are handled by our code before anything else can see them; and nothing about the photo is kept in places other scripts can reach. Ads are never placed next to a photo that may be sensitive. The privacy policy lists exactly what ad partners can and can't see, including the limits of these protections.
How accurate is it?
It's an estimate. X's real system uses private models, more signals than a single image (such as account history, reports and text) and human review, and it changes without notice. Our tests on held-out public data show the Accurate model catching most clearly explicit images while rarely flagging everyday photos. Swimwear, art, medical images and news photos are the hardest cases for any automated system, including X's, so treat a Borderline result as "could go either way." The Quick model is tuned to be cautious and flags more borderline photos than Accurate.
Tips if your photo is likely to be flagged
- Mark it as sensitive yourself. Self-labeling is what X asks for, and it protects your account from an account-wide label.
- Crop or reframe so the sensitive part isn't the focus, or pick a different shot for the post preview.
- For graphic news imagery, add context in the text and consider linking to the photo rather than embedding it.
- Never post anything X doesn't allow at all; a label won't fix that.
Frequently asked questions
Is Will It Flag? made by X or Twitter?
No. It's an independent project and is not affiliated with, endorsed by or connected to X Corp. "X" and "Twitter" are used only to describe what the tool estimates.
Does it work for videos or GIFs?
Not yet. It checks still images (JPG, PNG, WebP, and HEIC where your browser supports it). For a GIF, the first frame is checked.
Why does the first check take longer?
The checker needs a moment to get ready in your browser the first time. After that, checks take a fraction of a second to a couple of seconds depending on your device.
Quick or Accurate?
Quick is the default: it's lighter, gets ready faster and suits phones; it's a little more cautious. Accurate uses the larger CLIP model with the head built on X's design and gives the best estimate.
Do you keep my photos or results?
No. Nothing about your photo is sent anywhere. We only remember your Quick/Accurate preference in your browser.
Credits and licenses
OpenAI CLIP (MIT license). LAION CLIP-based detector and public embedding datasets (MIT / CC-BY 4.0). ONNX Runtime Web (MIT). Architecture follows the Apache-2.0 licensed code in X's "the-algorithm" repository. Will It Flag? is not affiliated with any of these projects.
Contact
Questions, feedback or a takedown request? Email contact@willitflag.com. Please don't email photos; we don't need them, and the tool never asks for them.